uvm_anon.c revision 1.16 
1/*	$NetBSD: uvm_anon.c,v 1.16 2001/03/10 22:46:47 chs Exp $	*/
2
3/*
4 *
5 * Copyright (c) 1997 Charles D. Cranor and Washington University.
6 * All rights reserved.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 *    notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 *    notice, this list of conditions and the following disclaimer in the
15 *    documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 *    must display the following acknowledgement:
18 *      This product includes software developed by Charles D. Cranor and
19 *      Washington University.
20 * 4. The name of the author may not be used to endorse or promote products
21 *    derived from this software without specific prior written permission.
22 *
23 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
24 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
25 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
26 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
27 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
28 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
32 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33 */
34
35/*
36 * uvm_anon.c: uvm anon ops
37 */
38
39#include "opt_uvmhist.h"
40
41#include <sys/param.h>
42#include <sys/systm.h>
43#include <sys/proc.h>
44#include <sys/malloc.h>
45#include <sys/pool.h>
46#include <sys/kernel.h>
47
48#include <uvm/uvm.h>
49#include <uvm/uvm_swap.h>
50
51/*
52 * anonblock_list: global list of anon blocks,
53 * locked by swap_syscall_lock (since we never remove
54 * anything from this list and we only add to it via swapctl(2)).
55 */
56
57struct uvm_anonblock {
58	LIST_ENTRY(uvm_anonblock) list;
59	int count;
60	struct vm_anon *anons;
61};
62static LIST_HEAD(anonlist, uvm_anonblock) anonblock_list;
63
64
65static boolean_t anon_pagein __P((struct vm_anon *));
66
67
68/*
69 * allocate anons
70 */
71void
72uvm_anon_init()
73{
74	int nanon = uvmexp.free - (uvmexp.free / 16); /* XXXCDC ??? */
75
76	simple_lock_init(&uvm.afreelock);
77	LIST_INIT(&anonblock_list);
78
79	/*
80	 * Allocate the initial anons.
81	 */
82	uvm_anon_add(nanon);
83}
84
85/*
86 * add some more anons to the free pool.  called when we add
87 * more swap space.
88 *
89 * => swap_syscall_lock should be held (protects anonblock_list).
90 */
91int
92uvm_anon_add(count)
93	int	count;
94{
95	struct uvm_anonblock *anonblock;
96	struct vm_anon *anon;
97	int lcv, needed;
98
99	simple_lock(&uvm.afreelock);
100	uvmexp.nanonneeded += count;
101	needed = uvmexp.nanonneeded - uvmexp.nanon;
102	simple_unlock(&uvm.afreelock);
103
104	if (needed <= 0) {
105		return 0;
106	}
107	anon = (void *)uvm_km_alloc(kernel_map, sizeof(*anon) * needed);
108	if (anon == NULL) {
109		simple_lock(&uvm.afreelock);
110		uvmexp.nanonneeded -= count;
111		simple_unlock(&uvm.afreelock);
112		return ENOMEM;
113	}
114	MALLOC(anonblock, void *, sizeof(*anonblock), M_UVMAMAP, M_WAITOK);
115
116	anonblock->count = needed;
117	anonblock->anons = anon;
118	LIST_INSERT_HEAD(&anonblock_list, anonblock, list);
119	memset(anon, 0, sizeof(*anon) * needed);
120
121	simple_lock(&uvm.afreelock);
122	uvmexp.nanon += needed;
123	uvmexp.nfreeanon += needed;
124	for (lcv = 0; lcv < needed; lcv++) {
125		simple_lock_init(&anon->an_lock);
126		anon[lcv].u.an_nxt = uvm.afree;
127		uvm.afree = &anon[lcv];
128		simple_lock_init(&uvm.afree->an_lock);
129	}
130	simple_unlock(&uvm.afreelock);
131	return 0;
132}
133
134/*
135 * remove anons from the free pool.
136 */
137void
138uvm_anon_remove(count)
139	int count;
140{
141	/*
142	 * we never actually free any anons, to avoid allocation overhead.
143	 * XXX someday we might want to try to free anons.
144	 */
145
146	simple_lock(&uvm.afreelock);
147	uvmexp.nanonneeded -= count;
148	simple_unlock(&uvm.afreelock);
149}
150
151/*
152 * allocate an anon
153 *
154 * => new anon is returned locked!
155 */
156struct vm_anon *
157uvm_analloc()
158{
159	struct vm_anon *a;
160
161	simple_lock(&uvm.afreelock);
162	a = uvm.afree;
163	if (a) {
164		uvm.afree = a->u.an_nxt;
165		uvmexp.nfreeanon--;
166		a->an_ref = 1;
167		a->an_swslot = 0;
168		a->u.an_page = NULL;		/* so we can free quickly */
169		LOCK_ASSERT(simple_lock_held(&a->an_lock) == 0);
170		simple_lock(&a->an_lock);
171	}
172	simple_unlock(&uvm.afreelock);
173	return(a);
174}
175
176/*
177 * uvm_anfree: free a single anon structure
178 *
179 * => caller must remove anon from its amap before calling (if it was in
180 *	an amap).
181 * => anon must be unlocked and have a zero reference count.
182 * => we may lock the pageq's.
183 */
184void
185uvm_anfree(anon)
186	struct vm_anon *anon;
187{
188	struct vm_page *pg;
189	UVMHIST_FUNC("uvm_anfree"); UVMHIST_CALLED(maphist);
190	UVMHIST_LOG(maphist,"(anon=0x%x)", anon, 0,0,0);
191
192	KASSERT(anon->an_ref == 0);
193	LOCK_ASSERT(simple_lock_held(&anon->an_lock) == 0);
194
195	/*
196	 * get page
197	 */
198
199	pg = anon->u.an_page;
200
201	/*
202	 * if there is a resident page and it is loaned, then anon may not
203	 * own it.   call out to uvm_anon_lockpage() to ensure the real owner
204 	 * of the page has been identified and locked.
205	 */
206
207	if (pg && pg->loan_count)
208		pg = uvm_anon_lockloanpg(anon);
209
210	/*
211	 * if we have a resident page, we must dispose of it before freeing
212	 * the anon.
213	 */
214
215	if (pg) {
216
217		/*
218		 * if the page is owned by a uobject (now locked), then we must
219		 * kill the loan on the page rather than free it.
220		 */
221
222		if (pg->uobject) {
223			uvm_lock_pageq();
224			KASSERT(pg->loan_count > 0);
225			pg->loan_count--;
226			pg->uanon = NULL;
227			uvm_unlock_pageq();
228			simple_unlock(&pg->uobject->vmobjlock);
229		} else {
230
231			/*
232			 * page has no uobject, so we must be the owner of it.
233			 *
234			 * if page is busy then we just mark it as released
235			 * (who ever has it busy must check for this when they
236			 * wake up).    if the page is not busy then we can
237			 * free it now.
238			 */
239
240			if ((pg->flags & PG_BUSY) != 0) {
241				/* tell them to dump it when done */
242				pg->flags |= PG_RELEASED;
243				UVMHIST_LOG(maphist,
244				    "  anon 0x%x, page 0x%x: BUSY (released!)",
245				    anon, pg, 0, 0);
246				return;
247			}
248			pmap_page_protect(pg, VM_PROT_NONE);
249			uvm_lock_pageq();	/* lock out pagedaemon */
250			uvm_pagefree(pg);	/* bye bye */
251			uvm_unlock_pageq();	/* free the daemon */
252			UVMHIST_LOG(maphist,"anon 0x%x, page 0x%x: freed now!",
253			    anon, pg, 0, 0);
254		}
255	}
256
257	/*
258	 * free any swap resources.
259	 */
260	uvm_anon_dropswap(anon);
261
262	/*
263	 * now that we've stripped the data areas from the anon, free the anon
264	 * itself!
265	 */
266	simple_lock(&uvm.afreelock);
267	anon->u.an_nxt = uvm.afree;
268	uvm.afree = anon;
269	uvmexp.nfreeanon++;
270	simple_unlock(&uvm.afreelock);
271	UVMHIST_LOG(maphist,"<- done!",0,0,0,0);
272}
273
274/*
275 * uvm_anon_dropswap:  release any swap resources from this anon.
276 *
277 * => anon must be locked or have a reference count of 0.
278 */
279void
280uvm_anon_dropswap(anon)
281	struct vm_anon *anon;
282{
283	UVMHIST_FUNC("uvm_anon_dropswap"); UVMHIST_CALLED(maphist);
284
285	if (anon->an_swslot == 0)
286		return;
287
288	UVMHIST_LOG(maphist,"freeing swap for anon %p, paged to swslot 0x%x",
289		    anon, anon->an_swslot, 0, 0);
290	uvm_swap_free(anon->an_swslot, 1);
291	anon->an_swslot = 0;
292
293	if (anon->u.an_page == NULL) {
294		/* this page is no longer only in swap. */
295		simple_lock(&uvm.swap_data_lock);
296		uvmexp.swpgonly--;
297		simple_unlock(&uvm.swap_data_lock);
298	}
299}
300
301/*
302 * uvm_anon_lockloanpg: given a locked anon, lock its resident page
303 *
304 * => anon is locked by caller
305 * => on return: anon is locked
306 *		 if there is a resident page:
307 *			if it has a uobject, it is locked by us
308 *			if it is ownerless, we take over as owner
309 *		 we return the resident page (it can change during
310 *		 this function)
311 * => note that the only time an anon has an ownerless resident page
312 *	is if the page was loaned from a uvm_object and the uvm_object
313 *	disowned it
314 * => this only needs to be called when you want to do an operation
315 *	on an anon's resident page and that page has a non-zero loan
316 *	count.
317 */
318struct vm_page *
319uvm_anon_lockloanpg(anon)
320	struct vm_anon *anon;
321{
322	struct vm_page *pg;
323	boolean_t locked = FALSE;
324
325	LOCK_ASSERT(simple_lock_held(&anon->an_lock));
326
327	/*
328	 * loop while we have a resident page that has a non-zero loan count.
329	 * if we successfully get our lock, we will "break" the loop.
330	 * note that the test for pg->loan_count is not protected -- this
331	 * may produce false positive results.   note that a false positive
332	 * result may cause us to do more work than we need to, but it will
333	 * not produce an incorrect result.
334	 */
335
336	while (((pg = anon->u.an_page) != NULL) && pg->loan_count != 0) {
337
338		/*
339		 * quickly check to see if the page has an object before
340		 * bothering to lock the page queues.   this may also produce
341		 * a false positive result, but that's ok because we do a real
342		 * check after that.
343		 *
344		 * XXX: quick check -- worth it?   need volatile?
345		 */
346
347		if (pg->uobject) {
348
349			uvm_lock_pageq();
350			if (pg->uobject) {	/* the "real" check */
351				locked =
352				    simple_lock_try(&pg->uobject->vmobjlock);
353			} else {
354				/* object disowned before we got PQ lock */
355				locked = TRUE;
356			}
357			uvm_unlock_pageq();
358
359			/*
360			 * if we didn't get a lock (try lock failed), then we
361			 * toggle our anon lock and try again
362			 */
363
364			if (!locked) {
365				simple_unlock(&anon->an_lock);
366
367				/*
368				 * someone locking the object has a chance to
369				 * lock us right now
370				 */
371
372				simple_lock(&anon->an_lock);
373				continue;
374			}
375		}
376
377		/*
378		 * if page is un-owned [i.e. the object dropped its ownership],
379		 * then we can take over as owner!
380		 */
381
382		if (pg->uobject == NULL && (pg->pqflags & PQ_ANON) == 0) {
383			uvm_lock_pageq();
384			pg->pqflags |= PQ_ANON;		/* take ownership... */
385			pg->loan_count--;	/* ... and drop our loan */
386			uvm_unlock_pageq();
387		}
388
389		/*
390		 * we did it!   break the loop
391		 */
392
393		break;
394	}
395	return(pg);
396}
397
398
399
400/*
401 * page in every anon that is paged out to a range of swslots.
402 *
403 * swap_syscall_lock should be held (protects anonblock_list).
404 */
405
406boolean_t
407anon_swap_off(startslot, endslot)
408	int startslot, endslot;
409{
410	struct uvm_anonblock *anonblock;
411
412	for (anonblock = LIST_FIRST(&anonblock_list);
413	     anonblock != NULL;
414	     anonblock = LIST_NEXT(anonblock, list)) {
415		int i;
416
417		/*
418		 * loop thru all the anons in the anonblock,
419		 * paging in where needed.
420		 */
421
422		for (i = 0; i < anonblock->count; i++) {
423			struct vm_anon *anon = &anonblock->anons[i];
424			int slot;
425
426			/*
427			 * lock anon to work on it.
428			 */
429
430			simple_lock(&anon->an_lock);
431
432			/*
433			 * is this anon's swap slot in range?
434			 */
435
436			slot = anon->an_swslot;
437			if (slot >= startslot && slot < endslot) {
438				boolean_t rv;
439
440				/*
441				 * yup, page it in.
442				 */
443
444				/* locked: anon */
445				rv = anon_pagein(anon);
446				/* unlocked: anon */
447
448				if (rv) {
449					return rv;
450				}
451			} else {
452
453				/*
454				 * nope, unlock and proceed.
455				 */
456
457				simple_unlock(&anon->an_lock);
458			}
459		}
460	}
461	return FALSE;
462}
463
464
465/*
466 * fetch an anon's page.
467 *
468 * => anon must be locked, and is unlocked upon return.
469 * => returns TRUE if pagein was aborted due to lack of memory.
470 */
471
472static boolean_t
473anon_pagein(anon)
474	struct vm_anon *anon;
475{
476	struct vm_page *pg;
477	struct uvm_object *uobj;
478	int rv;
479
480	/* locked: anon */
481	LOCK_ASSERT(simple_lock_held(&anon->an_lock));
482
483	rv = uvmfault_anonget(NULL, NULL, anon);
484
485	/*
486	 * if rv == 0, anon is still locked, else anon
487	 * is unlocked
488	 */
489
490	switch (rv) {
491	case 0:
492		break;
493
494	case EIO:
495	case ERESTART:
496
497		/*
498		 * nothing more to do on errors.
499		 * ERESTART can only mean that the anon was freed,
500		 * so again there's nothing to do.
501		 */
502
503		return FALSE;
504	}
505
506	/*
507	 * ok, we've got the page now.
508	 * mark it as dirty, clear its swslot and un-busy it.
509	 */
510
511	pg = anon->u.an_page;
512	uobj = pg->uobject;
513	uvm_swap_free(anon->an_swslot, 1);
514	anon->an_swslot = 0;
515	pg->flags &= ~(PG_CLEAN);
516
517	/*
518	 * deactivate the page (to put it on a page queue)
519	 */
520
521	pmap_clear_reference(pg);
522	uvm_lock_pageq();
523	uvm_pagedeactivate(pg);
524	uvm_unlock_pageq();
525
526	/*
527	 * unlock the anon and we're done.
528	 */
529
530	simple_unlock(&anon->an_lock);
531	if (uobj) {
532		simple_unlock(&uobj->vmobjlock);
533	}
534	return FALSE;
535}
536